Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Model Reduction in Control and Simulation:Algorithms and Applications
Peter Benner
Professur Mathematik in Industrie und TechnikFakultat fur Mathematik
Technische Universitat Chemnitz
Oxford UniversityComputational Mathematics and Applications Seminar
18 October 2007
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Current andFuture Work
References
Overview
1 IntroductionModel ReductionModel Reduction for Linear SystemsApplication Areas
2 Model ReductionGoalsModal TruncationPade ApproximationBalanced Truncation
3 ExamplesOptimal Control: Cooling of Steel ProfilesMicrothrusterButterfly GyroInterconnectReconstruction of the State
4 Current and Future Work
5 References
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Current andFuture Work
References
Joint work with
Enrique Quintana-Ortı, Gregorio Quintana-Ortı, RafaMayo, Jose Manuel Badıa, Alfredo Remon, SergioBarrachina (Universidad Jaume I de Castellon, Spain).
Ulrike Baur, Matthias Pester, Jens Saak( )
.
Viatcheslav Sokolov(
former).
Heike Faßbender (TU Braunschweig).
Infineon Technologies/Qimonda, IMTEK (U Freiburg),iwb (TU Munchen), . . .
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Introduction
Problem
Given a physical problem with dynamics described by the statesx ∈ Rn, where n is the dimension of the state space.
Because of redundancies, complexity, etc., we want to describe thedynamics of the system using a reduced number of states.
This is the task of model reduction (also: dimension reduction, orderreduction).
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Introduction
Problem
Given a physical problem with dynamics described by the statesx ∈ Rn, where n is the dimension of the state space.
Because of redundancies, complexity, etc., we want to describe thedynamics of the system using a reduced number of states.
This is the task of model reduction (also: dimension reduction, orderreduction).
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Introduction
Problem
Given a physical problem with dynamics described by the statesx ∈ Rn, where n is the dimension of the state space.
Because of redundancies, complexity, etc., we want to describe thedynamics of the system using a reduced number of states.
This is the task of model reduction (also: dimension reduction, orderreduction).
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Motivation: Image Compression by Truncated SVD
A digital image with nx × ny pixels can be represented as matrixX ∈ Rnx×ny , where xij contains color information of pixel (i , j).
Memory: 4 · nx · ny bytes.
Theorem: (Schmidt-Mirsky/Eckart-Young)
Best rank-r approximation to X ∈ Rnx×ny w.r.t. spectral norm:
X =∑r
j=1σjujv
Tj ,
where X = UΣV T is the singular value decomposition (SVD) of X .
The approximation error is ‖X − X‖2 = σr+1.
Idea for dimension reduction
Instead of X save u1, . . . , ur , σ1v1, . . . , σrvr . memory = 4r × (nx + ny ) bytes.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Motivation: Image Compression by Truncated SVD
A digital image with nx × ny pixels can be represented as matrixX ∈ Rnx×ny , where xij contains color information of pixel (i , j).
Memory: 4 · nx · ny bytes.
Theorem: (Schmidt-Mirsky/Eckart-Young)
Best rank-r approximation to X ∈ Rnx×ny w.r.t. spectral norm:
X =∑r
j=1σjujv
Tj ,
where X = UΣV T is the singular value decomposition (SVD) of X .
The approximation error is ‖X − X‖2 = σr+1.
Idea for dimension reduction
Instead of X save u1, . . . , ur , σ1v1, . . . , σrvr . memory = 4r × (nx + ny ) bytes.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Motivation: Image Compression by Truncated SVD
A digital image with nx × ny pixels can be represented as matrixX ∈ Rnx×ny , where xij contains color information of pixel (i , j).
Memory: 4 · nx · ny bytes.
Theorem: (Schmidt-Mirsky/Eckart-Young)
Best rank-r approximation to X ∈ Rnx×ny w.r.t. spectral norm:
X =∑r
j=1σjujv
Tj ,
where X = UΣV T is the singular value decomposition (SVD) of X .
The approximation error is ‖X − X‖2 = σr+1.
Idea for dimension reduction
Instead of X save u1, . . . , ur , σ1v1, . . . , σrvr . memory = 4r × (nx + ny ) bytes.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Example: Image Compression by Truncated SVD
Example: Clown
320× 200 pixel ≈ 256 kb
rank r = 50, ≈ 104 kb
rank r = 20, ≈ 42 kb
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Example: Image Compression by Truncated SVD
Example: Clown
320× 200 pixel ≈ 256 kb
rank r = 50, ≈ 104 kb
rank r = 20, ≈ 42 kb
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Example: Image Compression by Truncated SVD
Example: Clown
320× 200 pixel ≈ 256 kb
rank r = 50, ≈ 104 kb
rank r = 20, ≈ 42 kb
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Dimension Reduction via SVD
Example: Gatlinburg
Organizing committeeGatlinburg/Householder Meeting 1964:
James H. Wilkinson, Wallace Givens,
George Forsythe, Alston Householder,
Peter Henrici, Fritz L. Bauer.
640× 480 pixel, ≈ 1229 kb
rank r = 100, ≈ 448 kb
rank r = 50, ≈ 224 kb
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Dimension Reduction via SVD
Example: Gatlinburg
Organizing committeeGatlinburg/Householder Meeting 1964:
James H. Wilkinson, Wallace Givens,
George Forsythe, Alston Householder,
Peter Henrici, Fritz L. Bauer.
640× 480 pixel, ≈ 1229 kb
rank r = 100, ≈ 448 kb
rank r = 50, ≈ 224 kb
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Dimension Reduction via SVD
Example: Gatlinburg
Organizing committeeGatlinburg/Householder Meeting 1964:
James H. Wilkinson, Wallace Givens,
George Forsythe, Alston Householder,
Peter Henrici, Fritz L. Bauer.
640× 480 pixel, ≈ 1229 kb
rank r = 100, ≈ 448 kb
rank r = 50, ≈ 224 kb
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Background
Image data compression via SVD works, if the singular values decay(exponentially).
Singular Values of the Image Data Matrices
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
The Model Reduction Problem
Dynamical Systems
Σ :
x(t) = f (t, x(t), u(t)), x(t0) = x0,y(t) = g(t, x(t), u(t))
with
states x(t) ∈ Rn,
inputs u(t) ∈ Rm,
outputs y(t) ∈ Rp.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Model Reduction for Dynamical Systems
Original System
Σ :
x(t) = f (t, x(t), u(t)),y(t) = g(t, x(t), u(t)).
states x(t) ∈ Rn,
inputs u(t) ∈ Rm,
outputs y(t) ∈ Rp.
Reduced-Order System
bΣ :
˙x(t) = bf (t, x(t), u(t)),y(t) = bg(t, x(t), u(t)).
states x(t) ∈ Rr , r n
inputs u(t) ∈ Rm,
outputs y(t) ∈ Rp.
Goal:
‖y − y‖ < tolerance · ‖u‖ for all admissible input signals.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Model Reduction for Dynamical Systems
Original System
Σ :
x(t) = f (t, x(t), u(t)),y(t) = g(t, x(t), u(t)).
states x(t) ∈ Rn,
inputs u(t) ∈ Rm,
outputs y(t) ∈ Rp.
Reduced-Order System
bΣ :
˙x(t) = bf (t, x(t), u(t)),y(t) = bg(t, x(t), u(t)).
states x(t) ∈ Rr , r n
inputs u(t) ∈ Rm,
outputs y(t) ∈ Rp.
Goal:
‖y − y‖ < tolerance · ‖u‖ for all admissible input signals.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Model Reduction for Dynamical Systems
Original System
Σ :
x(t) = f (t, x(t), u(t)),y(t) = g(t, x(t), u(t)).
states x(t) ∈ Rn,
inputs u(t) ∈ Rm,
outputs y(t) ∈ Rp.
Reduced-Order System
bΣ :
˙x(t) = bf (t, x(t), u(t)),y(t) = bg(t, x(t), u(t)).
states x(t) ∈ Rr , r n
inputs u(t) ∈ Rm,
outputs y(t) ∈ Rp.
Goal:
‖y − y‖ < tolerance · ‖u‖ for all admissible input signals.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Model Reduction for Linear Systems
Linear, Time-Invariant (LTI) Systems
x(t) = f (t, x , u) = Ax + Bu, A ∈ Rn×n, B ∈ Rn×m,y(t) = g(t, x , u) = Cx + Du, C ∈ Rp×n, D ∈ Rp×m.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Model Reduction for Linear Systems
Linear, Time-Invariant (LTI) Systems
x(t) = f (t, x , u) = Ax + Bu, A ∈ Rn×n, B ∈ Rn×m,y(t) = g(t, x , u) = Cx + Du, C ∈ Rp×n, D ∈ Rp×m.
State-Space Description for I/O-Relation (D = 0)
S : u 7→ y , y(t) = (h ? u)(t) =
∫ ∞−∞
h(t − τ)u(τ) dτ,
where h(s) =
CeAsB if s > 00 if s ≤ 0
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Model Reduction for Linear Systems
Linear, Time-Invariant (LTI) Systems
x(t) = f (t, x , u) = Ax + Bu, A ∈ Rn×n, B ∈ Rn×m,y(t) = g(t, x , u) = Cx + Du, C ∈ Rp×n, D ∈ Rp×m.
State-Space Description for I/O-Relation (D = 0)
S : u 7→ y , y(t) = (h ? u)(t) =
∫ ∞−∞
h(t − τ)u(τ) dτ,
where h(s) =
CeAsB if s > 00 if s ≤ 0
Note: operator S not suitable for approximation as singular values arecontinuous; for model reduction use Hankel operator H.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Model Reduction for Linear Systems
Linear, Time-Invariant (LTI) Systems
x(t) = f (t, x , u) = Ax + Bu, A ∈ Rn×n, B ∈ Rn×m,y(t) = g(t, x , u) = Cx + Du, C ∈ Rp×n, D ∈ Rp×m.
State-Space Description for I/O-Relation (D = 0)
H : u− 7→ y+, y+(t) =∫ 0
−∞ CeA(t−τ)Bu(τ) dτ for all t > 0.
Note: operator S not suitable for approximation as singular values arecontinuous; for model reduction use Hankel operator H.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Model Reduction for Linear Systems
Linear, Time-Invariant (LTI) Systems
x(t) = f (t, x , u) = Ax + Bu, A ∈ Rn×n, B ∈ Rn×m,y(t) = g(t, x , u) = Cx + Du, C ∈ Rp×n, D ∈ Rp×m.
State-Space Description for I/O-Relation (D = 0)
H : u− 7→ y+, y+(t) =∫ 0
−∞ CeA(t−τ)Bu(τ) dτ for all t > 0.
Note: operator S not suitable for approximation as singular values arecontinuous; for model reduction use Hankel operator H.
H compact ⇒ H has discrete SVD Hankel singular values
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Model Reduction for Linear Systems
Linear, Time-Invariant (LTI) Systems
x(t) = f (t, x , u) = Ax + Bu, A ∈ Rn×n, B ∈ Rn×m,y(t) = g(t, x , u) = Cx + Du, C ∈ Rp×n, D ∈ Rp×m.
State-Space Description for I/O-Relation (D = 0)
H : u− 7→ y+, y+(t) =∫ 0
−∞ CeA(t−τ)Bu(τ) dτ for all t > 0.
H compact ⇒H has discrete SVD Hankel singular values
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Model Reduction for Linear Systems
Linear, Time-Invariant (LTI) Systems
x(t) = f (t, x , u) = Ax + Bu, A ∈ Rn×n, B ∈ Rn×m,y(t) = g(t, x , u) = Cx + Du, C ∈ Rp×n, D ∈ Rp×m.
State-Space Description for I/O-Relation (D = 0)
H : u− 7→ y+, y+(t) =∫ 0
−∞ CeA(t−τ)Bu(τ) dτ for all t > 0.
H compact ⇒ H has discrete SVD
⇒ Best approx. problem w.r.t. 2-induced operator norm(Hankel norm) well-posed.
⇒ solution: Adamjan-Arov-Krein (AAK Theory, 1971/78).
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Model Reduction for Linear Systems
Linear, Time-Invariant (LTI) Systems
x(t) = f (t, x , u) = Ax + Bu, A ∈ Rn×n, B ∈ Rn×m,y(t) = g(t, x , u) = Cx + Du, C ∈ Rp×n, D ∈ Rp×m.
State-Space Description for I/O-Relation (D = 0)
H : u− 7→ y+, y+(t) =∫ 0
−∞ CeA(t−τ)Bu(τ) dτ for all t > 0.
H compact ⇒ H has discrete SVD
⇒ Best approx. problem w.r.t. 2-induced operator norm(Hankel norm) well-posed.
⇒ solution: Adamjan-Arov-Krein (AAK Theory, 1971/78).
But: computationally unfeasible for large-scale systems.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Linear Systems in Frequency Domain
Linear, Time-Invariant (LTI) Systems
f (t, x , u) = Ax + Bu, A ∈ Rn×n, B ∈ Rn×m,g(t, x , u) = Cx + Du, C ∈ Rp×n, D ∈ Rp×m.
Laplace Transformation / Frequency Domain
Application of Laplace transformation (x(t) 7→ x(s), x(t) 7→ sx(s))to linear system with x(0) = 0:
sx(s) = Ax(s) + Bu(s), y(s) = Bx(s) + Du(s),
yields I/O-relation in frequency domain:
y(s) =(
C (sIn − A)−1B + D︸ ︷︷ ︸=:G(s)
)u(s)
G is the transfer function of Σ.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Linear Systems in Frequency Domain
Linear, Time-Invariant (LTI) Systems
f (t, x , u) = Ax + Bu, A ∈ Rn×n, B ∈ Rn×m,g(t, x , u) = Cx + Du, C ∈ Rp×n, D ∈ Rp×m.
Laplace Transformation / Frequency Domain
Application of Laplace transformation (x(t) 7→ x(s), x(t) 7→ sx(s))to linear system with x(0) = 0:
sx(s) = Ax(s) + Bu(s), y(s) = Bx(s) + Du(s),
yields I/O-relation in frequency domain:
y(s) =(
C (sIn − A)−1B + D︸ ︷︷ ︸=:G(s)
)u(s)
G is the transfer function of Σ.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Linear Systems in Frequency Domain
Linear, Time-Invariant (LTI) Systems
f (t, x , u) = Ax + Bu, A ∈ Rn×n, B ∈ Rn×m,g(t, x , u) = Cx + Du, C ∈ Rp×n, D ∈ Rp×m.
Laplace Transformation / Frequency Domain
Application of Laplace transformation (x(t) 7→ x(s), x(t) 7→ sx(s))to linear system with x(0) = 0:
sx(s) = Ax(s) + Bu(s), y(s) = Bx(s) + Du(s),
yields I/O-relation in frequency domain:
y(s) =(
C (sIn − A)−1B + D︸ ︷︷ ︸=:G(s)
)u(s)
G is the transfer function of Σ.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Model Reduction for Linear Systems
Problem
Approximate the dynamical system
x = Ax + Bu, A ∈ Rn×n, B ∈ Rn×m,y = Cx + Du, C ∈ Rp×n, D ∈ Rp×m,
by reduced-order system
˙x = Ax + Bu, A ∈ Rr×r , B ∈ Rr×m,
y = C x + Du, C ∈ Rp×r , D ∈ Rp×m,
of order r n, such that
‖y − y‖ = ‖Gu − Gu‖ ≤ ‖G − G‖‖u‖ < tolerance · ‖u‖.
=⇒ Approximation problem: minorder (G)≤r ‖G − G‖.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Model Reduction for Linear Systems
Problem
Approximate the dynamical system
x = Ax + Bu, A ∈ Rn×n, B ∈ Rn×m,y = Cx + Du, C ∈ Rp×n, D ∈ Rp×m,
by reduced-order system
˙x = Ax + Bu, A ∈ Rr×r , B ∈ Rr×m,
y = C x + Du, C ∈ Rp×r , D ∈ Rp×m,
of order r n, such that
‖y − y‖ = ‖Gu − Gu‖ ≤ ‖G − G‖‖u‖ < tolerance · ‖u‖.
=⇒ Approximation problem: minorder (G)≤r ‖G − G‖.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Application Areas(Optimal) Control
Feedback Controllers
A feedback controller (dynamiccompensator) is a linear system oforder N, where
input = output of plant,
output = input of plant.
Modern (LQG-/H2-/H∞-) controldesign: N ≥ n
⇒ reduce order of original system.
Real-time control is only possible with controllers of low complexity.
Experience tells us: the more complex, the more fragile.
Modern feedback control of systems governed by PDEs impossibledue to large scale of systems arising from FE discretization.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Application Areas(Optimal) Control
Feedback Controllers
A feedback controller (dynamiccompensator) is a linear system oforder N, where
input = output of plant,
output = input of plant.
Modern (LQG-/H2-/H∞-) controldesign: N ≥ n
⇒ reduce order of original system.
Real-time control is only possible with controllers of low complexity.
Experience tells us: the more complex, the more fragile.
Modern feedback control of systems governed by PDEs impossibledue to large scale of systems arising from FE discretization.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Application Areas(Optimal) Control
Feedback Controllers
A feedback controller (dynamiccompensator) is a linear system oforder N, where
input = output of plant,
output = input of plant.
Modern (LQG-/H2-/H∞-) controldesign: N ≥ n
⇒ reduce order of original system.
Real-time control is only possible with controllers of low complexity.
Experience tells us: the more complex, the more fragile.
Modern feedback control of systems governed by PDEs impossibledue to large scale of systems arising from FE discretization.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Application Areas(Optimal) Control
Feedback Controllers
A feedback controller (dynamiccompensator) is a linear system oforder N, where
input = output of plant,
output = input of plant.
Modern (LQG-/H2-/H∞-) controldesign: N ≥ n
⇒ reduce order of original system.
Real-time control is only possible with controllers of low complexity.
Experience tells us: the more complex, the more fragile.
Modern feedback control of systems governed by PDEs impossibledue to large scale of systems arising from FE discretization.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Application Areas(Optimal) Control
Feedback Controllers
A feedback controller (dynamiccompensator) is a linear system oforder N, where
input = output of plant,
output = input of plant.
Modern (LQG-/H2-/H∞-) controldesign: N ≥ n
⇒ reduce order of original system.
Real-time control is only possible with controllers of low complexity.
Experience tells us: the more complex, the more fragile.
Modern feedback control of systems governed by PDEs impossibledue to large scale of systems arising from FE discretization.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Application AreasMicro Electronics
Progressive miniaturization: Moore’s Law states that thenumber of on-chip transistors doubles each 12 (now: 18) months.
Verification of VLSI/ULSI chip design requires high number ofsimulations for different input signals.
Increase in packing density requires modeling of interconncet toensure that thermic/electro-magnetic effects do not disturbsignal transmission.
Linear systems in micro electronics occur through modifiednodal analysis (MNA) for RLC networks, e.g., when
– decoupling large linear subcircuits,– modeling transmission lines,– modeling pin packages in VLSI chips,– modeling circuit elements described by Maxwell’s equation
using partial element equivalent circuits (PEEC).
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Application AreasMicro Electronics
Progressive miniaturization: Moore’s Law states that thenumber of on-chip transistors doubles each 12 (now: 18) months.
Verification of VLSI/ULSI chip design requires high number ofsimulations for different input signals.
Increase in packing density requires modeling of interconncet toensure that thermic/electro-magnetic effects do not disturbsignal transmission.
Linear systems in micro electronics occur through modifiednodal analysis (MNA) for RLC networks, e.g., when
– decoupling large linear subcircuits,– modeling transmission lines,– modeling pin packages in VLSI chips,– modeling circuit elements described by Maxwell’s equation
using partial element equivalent circuits (PEEC).
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Application AreasMicro Electronics
Progressive miniaturization: Moore’s Law states that thenumber of on-chip transistors doubles each 12 (now: 18) months.
Verification of VLSI/ULSI chip design requires high number ofsimulations for different input signals.
Increase in packing density requires modeling of interconncet toensure that thermic/electro-magnetic effects do not disturbsignal transmission.
Linear systems in micro electronics occur through modifiednodal analysis (MNA) for RLC networks, e.g., when
– decoupling large linear subcircuits,– modeling transmission lines,– modeling pin packages in VLSI chips,– modeling circuit elements described by Maxwell’s equation
using partial element equivalent circuits (PEEC).
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Application AreasMicro Electronics
Progressive miniaturization: Moore’s Law states that thenumber of on-chip transistors doubles each 12 (now: 18) months.
Verification of VLSI/ULSI chip design requires high number ofsimulations for different input signals.
Increase in packing density requires modeling of interconncet toensure that thermic/electro-magnetic effects do not disturbsignal transmission.
Linear systems in micro electronics occur through modifiednodal analysis (MNA) for RLC networks, e.g., when
– decoupling large linear subcircuits,– modeling transmission lines,– modeling pin packages in VLSI chips,– modeling circuit elements described by Maxwell’s equation
using partial element equivalent circuits (PEEC).
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Application AreasMicro Electronics: Example for Miniaturization
Intel 4004 (1971)
1 layer, 10µ technology,
2,300 transistors,
64 kHz clock speed.
Intel Pentium IV (2001)
7 layers, 0.18µ technology,
42,000,000 transistors,
2 GHz clock speed,
2km of interconnect.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Application AreasMEMS/Microsystems
Typical problem in MEMS simulation:coupling of different models (thermic, structural, electric,electro-magnetic) during simulation.
Problems and Challenges:
Reduce simulation times by replacing sub-systems with theirreduced-order models.
Stability properties of coupled system may deteriorate throughmodel reduction even when stable sub-systems are replaced bystable reduced-order models.
Multi-scale phenomena.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Application AreasMEMS/Microsystems
Typical problem in MEMS simulation:coupling of different models (thermic, structural, electric,electro-magnetic) during simulation.
Problems and Challenges:
Reduce simulation times by replacing sub-systems with theirreduced-order models.
Stability properties of coupled system may deteriorate throughmodel reduction even when stable sub-systems are replaced bystable reduced-order models.
Multi-scale phenomena.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Application AreasMEMS/Microsystems
Typical problem in MEMS simulation:coupling of different models (thermic, structural, electric,electro-magnetic) during simulation.
Problems and Challenges:
Reduce simulation times by replacing sub-systems with theirreduced-order models.
Stability properties of coupled system may deteriorate throughmodel reduction even when stable sub-systems are replaced bystable reduced-order models.
Multi-scale phenomena.
Model Reduction
Peter Benner
Introduction
Model Reduction
Linear Systems
ApplicationAreas
Model Reduction
Examples
Current andFuture Work
References
Application AreasMEMS/Microsystems
Typical problem in MEMS simulation:coupling of different models (thermic, structural, electric,electro-magnetic) during simulation.
Problems and Challenges:
Reduce simulation times by replacing sub-systems with theirreduced-order models.
Stability properties of coupled system may deteriorate throughmodel reduction even when stable sub-systems are replaced bystable reduced-order models.
Multi-scale phenomena.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Model ReductionGoals
Automatic generation of compact models.
Satisfy desired error tolerance for all admissible input signals,i.e., want
‖y − y‖ < tolerance · ‖u‖ ∀u ∈ L2(R,Rm).
=⇒ Need computable error bound/estimate!
Preserve physical properties:
– stability (poles of G in C−),– minimum phase (zeroes of G in C−),– passivity (“system does not generate energy”).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Model ReductionGoals
Automatic generation of compact models.
Satisfy desired error tolerance for all admissible input signals,i.e., want
‖y − y‖ < tolerance · ‖u‖ ∀u ∈ L2(R,Rm).
=⇒ Need computable error bound/estimate!
Preserve physical properties:
– stability (poles of G in C−),– minimum phase (zeroes of G in C−),– passivity (“system does not generate energy”).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Model ReductionGoals
Automatic generation of compact models.
Satisfy desired error tolerance for all admissible input signals,i.e., want
‖y − y‖ < tolerance · ‖u‖ ∀u ∈ L2(R,Rm).
=⇒ Need computable error bound/estimate!
Preserve physical properties:
– stability (poles of G in C−),– minimum phase (zeroes of G in C−),– passivity (“system does not generate energy”).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Model ReductionGoals
Automatic generation of compact models.
Satisfy desired error tolerance for all admissible input signals,i.e., want
‖y − y‖ < tolerance · ‖u‖ ∀u ∈ L2(R,Rm).
=⇒ Need computable error bound/estimate!
Preserve physical properties:
– stability (poles of G in C−),– minimum phase (zeroes of G in C−),– passivity (“system does not generate energy”).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Model ReductionGoals
Automatic generation of compact models.
Satisfy desired error tolerance for all admissible input signals,i.e., want
‖y − y‖ < tolerance · ‖u‖ ∀u ∈ L2(R,Rm).
=⇒ Need computable error bound/estimate!
Preserve physical properties:
– stability (poles of G in C−),– minimum phase (zeroes of G in C−),– passivity (“system does not generate energy”).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Model ReductionGoals
Automatic generation of compact models.
Satisfy desired error tolerance for all admissible input signals,i.e., want
‖y − y‖ < tolerance · ‖u‖ ∀u ∈ L2(R,Rm).
=⇒ Need computable error bound/estimate!
Preserve physical properties:
– stability (poles of G in C−),– minimum phase (zeroes of G in C−),– passivity (“system does not generate energy”).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Model Reduction
1 Modal Truncation
2 Pade-Approximation and Krylov Subspace Methods
3 Balanced Truncation
4 many more. . .
Joint feature of many methods: Galerkin or Petrov-Galerkin-typeprojection of state-space onto low-dimensional subspace V along W:assume x ≈ VW T x =: x , where
range (V ) = V, range (W ) =W, W TV = Ir .
Then, with x = W T x , we obtain x ≈ V x and
‖x − x‖ = ‖x − V x‖.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Model Reduction
1 Modal Truncation
2 Pade-Approximation and Krylov Subspace Methods
3 Balanced Truncation
4 many more. . .
Joint feature of many methods: Galerkin or Petrov-Galerkin-typeprojection of state-space onto low-dimensional subspace V along W:assume x ≈ VW T x =: x , where
range (V ) = V, range (W ) =W, W TV = Ir .
Then, with x = W T x , we obtain x ≈ V x and
‖x − x‖ = ‖x − V x‖.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Modal Truncation
Idea:
Project state-space onto A-invariant subspace V, where
V = span(v1, . . . , vr ),
vk = eigenvectors corresp. to “dominant” modes ≡ eigenvalues of A.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Modal Truncation
Idea:
Project state-space onto A-invariant subspace V, where
V = span(v1, . . . , vr ),
vk = eigenvectors corresp. to “dominant” modes ≡ eigenvalues of A.
Properties:
Simple computation for large-scale systems, using, e.g., Krylovsubspace methods (Lanczos, Arnoldi), Jacobi-Davidson method.
Error bound:
‖G − G‖∞ ≤ cond2 (T ) ‖C2‖2‖B2‖21
minλ∈Λ (A2) |Re(λ)|,
where T−1AT = diag(A1,A2).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Modal Truncation
Idea:
Project state-space onto A-invariant subspace V, where
V = span(v1, . . . , vr ),
vk = eigenvectors corresp. to “dominant” modes ≡ eigenvalues of A.
Difficulties:
Eigenvalues contain only limited system information.
Dominance measures are difficult to compute.(Litz 1979: use Jordan canoncial form; otherwise merelyheuristic criteria.Rommes 2007: dominant pole algorithm (two-sided RQI).)
Error bound not computable for really large-scale probems.
New direction: AMLS (automated multilevel substructuring)[Benninghof/Lehoucq ’04, Elssel/Voß ’05, Blomeling ’06].
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Pade Approximation
Idea:
ConsiderEx = Ax + Bu, y = Cx
with rational transfer function G (s) = C (sE − A)−1B.
For s0 6∈ Λ (A,E ):
G (s) = m0 + m1(s − s0) + m2(s − s0)2 + . . .
As reduced-order model use rth Pade approximant G to G :
G (s) = G (s) +O((s − s0)2r ),
i.e., mj = mj for j = 0, . . . , 2r − 1
moment matching if s0 <∞,
partial realization if s0 =∞.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Pade Approximation
Idea:
ConsiderEx = Ax + Bu, y = Cx
with rational transfer function G (s) = C (sE − A)−1B.
For s0 6∈ Λ (A,E ):
G (s) = m0 + m1(s − s0) + m2(s − s0)2 + . . .
As reduced-order model use rth Pade approximant G to G :
G (s) = G (s) +O((s − s0)2r ),
i.e., mj = mj for j = 0, . . . , 2r − 1
moment matching if s0 <∞,
partial realization if s0 =∞.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Pade Approximation
Idea:
ConsiderEx = Ax + Bu, y = Cx
with rational transfer function G (s) = C (sE − A)−1B.
For s0 6∈ Λ (A,E ):
G (s) = m0 + m1(s − s0) + m2(s − s0)2 + . . .
As reduced-order model use rth Pade approximant G to G :
G (s) = G (s) +O((s − s0)2r ),
i.e., mj = mj for j = 0, . . . , 2r − 1
moment matching if s0 <∞,
partial realization if s0 =∞.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Pade Approximation
Pade-via-Lanczos Method (PVL)
Moments need not be computed explicitly; moment matching isequivalent to projecting state-space onto
V = span(B, AB, . . . , Ar−1B) = K(A, B, r)
(where A = (s0E − A)−1E , B = (s0E − A)−1B) along
W = span(CH , AHCH , . . . , (AH)r−1CH) = K(AH ,CH , r).
Computation via unsymmetric Lanczos method, yields systemmatrices of reduced-order model as by-product.
PVL applies w/o changes for singular E if s0 6∈ Λ (A,E ):– for s0 6=∞: Gallivan/Grimme/Van Dooren 1994,
Freund/Feldmann 1996, Grimme 1997
– for s0 =∞: B./Sokolov 2005
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Pade Approximation
Pade-via-Lanczos Method (PVL)
Moments need not be computed explicitly; moment matching isequivalent to projecting state-space onto
V = span(B, AB, . . . , Ar−1B) = K(A, B, r)
(where A = (s0E − A)−1E , B = (s0E − A)−1B) along
W = span(CH , AHCH , . . . , (AH)r−1CH) = K(AH ,CH , r).
Computation via unsymmetric Lanczos method, yields systemmatrices of reduced-order model as by-product.
PVL applies w/o changes for singular E if s0 6∈ Λ (A,E ):– for s0 6=∞: Gallivan/Grimme/Van Dooren 1994,
Freund/Feldmann 1996, Grimme 1997
– for s0 =∞: B./Sokolov 2005
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Pade Approximation
Pade-via-Lanczos Method (PVL)
Moments need not be computed explicitly; moment matching isequivalent to projecting state-space onto
V = span(B, AB, . . . , Ar−1B) = K(A, B, r)
(where A = (s0E − A)−1E , B = (s0E − A)−1B) along
W = span(CH , AHCH , . . . , (AH)r−1CH) = K(AH ,CH , r).
Computation via unsymmetric Lanczos method, yields systemmatrices of reduced-order model as by-product.
PVL applies w/o changes for singular E if s0 6∈ Λ (A,E ):– for s0 6=∞: Gallivan/Grimme/Van Dooren 1994,
Freund/Feldmann 1996, Grimme 1997
– for s0 =∞: B./Sokolov 2005
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Pade Approximation
Pade-via-Lanczos Method (PVL)
Partial realization for descriptor systems: [B./Sokolov, SCL, 2006]
For nonsingular E and s0 =∞:
moments = Markov parameters = C (E−1A)jE−1B, j = 0, 1, . . .
Question: for E singular, what is the correct generalized inverse here?
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Pade Approximation
Pade-via-Lanczos Method (PVL)
Partial realization for descriptor systems: [B./Sokolov, SCL, 2006]
For nonsingular E and s0 =∞:
moments = Markov parameters = C (E−1A)jE−1B, j = 0, 1, . . .
Question: for E singular, what is the correct generalized inverse here?
Answer: 2-inverse E2 = Q
[Inf
00 0
]P, where
sPEQ − PAQ = s
[Inf
00 N
]−[
J 00 In∞
],
is the Weierstraß canonical form (WCF) of sE − A.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Pade Approximation
Pade-via-Lanczos Method (PVL)
Partial realization for descriptor systems: [B./Sokolov, SCL, 2006]
For nonsingular E and s0 =∞:
moments = Markov parameters = C (E−1A)jE−1B, j = 0, 1, . . .
Question: for E singular, what is the correct generalized inverse here?
Answer: 2-inverse E2 = Q
[Inf
00 0
]P, where
sPEQ − PAQ = s
[Inf
00 N
]−[
J 00 In∞
],
is the Weierstraß canonical form (WCF) of sE − A.
P,Q can be computed w/o WCF in many applications.
Numerically, use Lanczos applied to E2A,E2B,C.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Pade Approximation
Pade-via-Lanczos Method (PVL)
Difficulties:
No computable error estimates/bounds for ‖y − y‖2.
Mostly heuristic criteria for choice of expansion points.Optimal choice for second-order systems with proportional/Rayleigh
damping (Beattie/Gugercin 2005).
Good approximation quality only locally.
Preservation of physical properties only in very special cases;usually requires post processing which (partially) destroysmoment matching properties.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Pade Approximation
Pade-via-Lanczos Method (PVL)
Difficulties:
No computable error estimates/bounds for ‖y − y‖2.
Mostly heuristic criteria for choice of expansion points.Optimal choice for second-order systems with proportional/Rayleigh
damping (Beattie/Gugercin 2005).
Good approximation quality only locally.
Preservation of physical properties only in very special cases;usually requires post processing which (partially) destroysmoment matching properties.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Pade Approximation
Pade-via-Lanczos Method (PVL)
Difficulties:
No computable error estimates/bounds for ‖y − y‖2.
Mostly heuristic criteria for choice of expansion points.Optimal choice for second-order systems with proportional/Rayleigh
damping (Beattie/Gugercin 2005).
Good approximation quality only locally.
Preservation of physical properties only in very special cases;usually requires post processing which (partially) destroysmoment matching properties.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Pade Approximation
Pade-via-Lanczos Method (PVL)
Difficulties:
No computable error estimates/bounds for ‖y − y‖2.
Mostly heuristic criteria for choice of expansion points.Optimal choice for second-order systems with proportional/Rayleigh
damping (Beattie/Gugercin 2005).
Good approximation quality only locally.
Preservation of physical properties only in very special cases;usually requires post processing which (partially) destroysmoment matching properties.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Pade Approximation
Pade-via-Lanczos Method (PVL)
Difficulties:
No computable error estimates/bounds for ‖y − y‖2.
Mostly heuristic criteria for choice of expansion points.Optimal choice for second-order systems with proportional/Rayleigh
damping (Beattie/Gugercin 2005).
Good approximation quality only locally.
Preservation of physical properties only in very special cases;usually requires post processing which (partially) destroysmoment matching properties.
New direction: moment matching yields rational interpolation ofG (j)(s) for j = 0, . . . , 2r − 1 at s = s0.Instead: use rational (Hermite) interpolation at sj , j = 0, . . . , r .
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Pade Approximation
Pade-via-Lanczos Method (PVL)
Difficulties:
No computable error estimates/bounds for ‖y − y‖2.
Mostly heuristic criteria for choice of expansion points.Optimal choice for second-order systems with proportional/Rayleigh
damping (Beattie/Gugercin 2005).
Good approximation quality only locally.
Preservation of physical properties only in very special cases;usually requires post processing which (partially) destroysmoment matching properties.
New direction: moment matching yields rational interpolation ofG (j)(s) for j = 0, . . . , 2r − 1 at s = s0.Instead: use rational (Hermite) interpolation at sj , j = 0, . . . , r .Current work: where to put the sj?
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Idea:
A system Σ, realized by (A,B,C ,D), is called balanced, ifsolutions P,Q of the Lyapunov equations
AP + PAT + BBT = 0, ATQ + QA + CTC = 0,
satisfy: P = Q = diag(σ1, . . . , σn) with σ1 ≥ σ2 ≥ . . . ≥ σn > 0.
σ1, . . . , σn are the Hankel singular values (HSVs) of Σ.
Compute balanced realization of the system via state-spacetransformation
T : (A,B,C ,D) 7→ (TAT−1,TB,CT−1,D)
=
„»A11 A12
A21 A22
–,
»B1
B2
–,ˆ
C1 C2
˜,D
«Truncation (A, B, C , D) = (A11,B1,C1,D).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Idea:
A system Σ, realized by (A,B,C ,D), is called balanced, ifsolutions P,Q of the Lyapunov equations
AP + PAT + BBT = 0, ATQ + QA + CTC = 0,
satisfy: P = Q = diag(σ1, . . . , σn) with σ1 ≥ σ2 ≥ . . . ≥ σn > 0.
σ1, . . . , σn are the Hankel singular values (HSVs) of Σ.
Compute balanced realization of the system via state-spacetransformation
T : (A,B,C ,D) 7→ (TAT−1,TB,CT−1,D)
=
„»A11 A12
A21 A22
–,
»B1
B2
–,ˆ
C1 C2
˜,D
«Truncation (A, B, C , D) = (A11,B1,C1,D).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Idea:
A system Σ, realized by (A,B,C ,D), is called balanced, ifsolutions P,Q of the Lyapunov equations
AP + PAT + BBT = 0, ATQ + QA + CTC = 0,
satisfy: P = Q = diag(σ1, . . . , σn) with σ1 ≥ σ2 ≥ . . . ≥ σn > 0.
σ1, . . . , σn are the Hankel singular values (HSVs) of Σ.
Compute balanced realization of the system via state-spacetransformation
T : (A,B,C ,D) 7→ (TAT−1,TB,CT−1,D)
=
„»A11 A12
A21 A22
–,
»B1
B2
–,ˆ
C1 C2
˜,D
«Truncation (A, B, C , D) = (A11,B1,C1,D).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Idea:
A system Σ, realized by (A,B,C ,D), is called balanced, ifsolutions P,Q of the Lyapunov equations
AP + PAT + BBT = 0, ATQ + QA + CTC = 0,
satisfy: P = Q = diag(σ1, . . . , σn) with σ1 ≥ σ2 ≥ . . . ≥ σn > 0.
σ1, . . . , σn are the Hankel singular values (HSVs) of Σ.
Compute balanced realization of the system via state-spacetransformation
T : (A,B,C ,D) 7→ (TAT−1,TB,CT−1,D)
=
„»A11 A12
A21 A22
–,
»B1
B2
–,ˆ
C1 C2
˜,D
«Truncation (A, B, C , D) = (A11,B1,C1,D).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Motivation:
HSV are system invariants: they are preserved under T and determinethe energy transfer given by the Hankel map
H : L2(−∞, 0) 7→ L2(0,∞) : u− 7→ y+.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Motivation:
HSV are system invariants: they are preserved under T and determinethe energy transfer given by the Hankel map
H : L2(−∞, 0) 7→ L2(0,∞) : u− 7→ y+.
In balanced coordinates . . . energy transfer from u− to y+:
E := supu∈L2(−∞,0]
x(0)=x0
∞∫0
y(t)T y(t) dt
0∫−∞
u(t)Tu(t) dt
=1
‖x0‖2
n∑j=1
σ2j x
20,j
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Motivation:
HSV are system invariants: they are preserved under T and determinethe energy transfer given by the Hankel map
H : L2(−∞, 0) 7→ L2(0,∞) : u− 7→ y+.
In balanced coordinates . . . energy transfer from u− to y+:
E := supu∈L2(−∞,0]
x(0)=x0
∞∫0
y(t)T y(t) dt
0∫−∞
u(t)Tu(t) dt
=1
‖x0‖2
n∑j=1
σ2j x
20,j
=⇒ Truncate states corresponding to “small” HSVs=⇒ complete analogy to best approximation via SVD!
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Implementation: SR Method
1 Compute (Cholesky) factors of the solutions of the Lyapunovequations,
P = STS , Q = RTR.
2 Compute SVD
SRT = [ U1, U2 ]
[Σ1
Σ2
] [V T
1
V T2
].
3 SetW = RTV1Σ
−1/21 , V = STU1Σ
−1/21 .
4 Reduced model is (W TAV ,W TB,CV ,D).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Implementation: SR Method
1 Compute (Cholesky) factors of the solutions of the Lyapunovequations,
P = STS , Q = RTR.
2 Compute SVD
SRT = [ U1, U2 ]
[Σ1
Σ2
] [V T
1
V T2
].
3 SetW = RTV1Σ
−1/21 , V = STU1Σ
−1/21 .
4 Reduced model is (W TAV ,W TB,CV ,D).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Implementation: SR Method
1 Compute (Cholesky) factors of the solutions of the Lyapunovequations,
P = STS , Q = RTR.
2 Compute SVD
SRT = [ U1, U2 ]
[Σ1
Σ2
] [V T
1
V T2
].
3 SetW = RTV1Σ
−1/21 , V = STU1Σ
−1/21 .
4 Reduced model is (W TAV ,W TB,CV ,D).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Properties:
Reduced-order model is stable with HSVs σ1, . . . , σr .
Adaptive choice of r via computable error bound:
‖y − y‖2 ≤(
2∑n
k=r+1σk
)‖u‖2.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Properties:
Reduced-order model is stable with HSVs σ1, . . . , σr .
Adaptive choice of r via computable error bound:
‖y − y‖2 ≤(
2∑n
k=r+1σk
)‖u‖2.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Properties:
General misconception: complexity O(n3) – true for severalimplementations! (e.g., Matlab, SLICOT, MorLAB).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Properties:
General misconception: complexity O(n3) – true for severalimplementations! (e.g., Matlab, SLICOT, MorLAB).
New algorithmic ideas from numerical linear algebra:
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Properties:
General misconception: complexity O(n3) – true for severalimplementations! (e.g., Matlab, SLICOT, MorLAB).
New algorithmic ideas from numerical linear algebra:
– Instead of Gramians P,Qcompute S ,R ∈ Rn×k ,k n, such that
P ≈ SST , Q ≈ RRT .
– Compute S ,R withproblem-specific Lyapunovsolvers of “low” complexitydirectly.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Properties:
General misconception: complexity O(n3) – true for severalimplementations! (e.g., Matlab, SLICOT, MorLAB).
New algorithmic ideas from numerical linear algebra:
Parallelization:
– Efficient parallel algorithms based on matrix sign function.
– Complexity O(n3/q) on q-processor machine.
– Software library PLiCMR with WebComputing interface.
(B./Quintana-Ortı/Quintana-Ortı since 1999)
Formatted Arithmetic:
For special problems from PDE control use implementation based onhierarchical matrices and matrix sign function method (Baur/B.),complexity O(n log2(n)r2).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Properties:
General misconception: complexity O(n3) – true for severalimplementations! (e.g., Matlab, SLICOT, MorLAB).
New algorithmic ideas from numerical linear algebra:
Parallelization:
– Efficient parallel algorithms based on matrix sign function.
– Complexity O(n3/q) on q-processor machine.
– Software library PLiCMR with WebComputing interface.
(B./Quintana-Ortı/Quintana-Ortı since 1999)
Formatted Arithmetic:
For special problems from PDE control use implementation based onhierarchical matrices and matrix sign function method (Baur/B.),complexity O(n log2(n)r2).
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation
Properties:
General misconception: complexity O(n3) – true for severalimplementations! (e.g., Matlab, SLICOT, MorLAB).
New algorithmic ideas from numerical linear algebra:
Sparse Balanced Truncation:
– Sparse implementation using sparse Lyapunov solver(ADI+MUMPS/SuperLU).
– Complexity O(n(k2 + r2)).
– Software:
+ Matlab toolbox LyaPack (Penzl 1999),+ Software library SpaRed with WebComputing interface.
(Badıa/B./Quintana-Ortı/Quintana-Ortı since 2003)
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation for Unstable Systems
For A stable, Gramians are defined by
P =
∫ ∞0
eAtBBT eAT t dt, Q =
∫ ∞0
eAT tCTCeAt dt.
For unstable A, integrals diverge!
Frequency-domain definition of Gramians
P := 12π
∫∞−∞(jω − A)−1BBT (jω − A)−H dω,
Q := 12π
∫∞−∞(jω − A)−HCTC (jω − A)−1 dω.
Well-defined if Λ (A) ∩ ıR = ∅!For stable A, definitions coincide.
Balancing/balanced truncation can be based on P, Q.Moreover, BT error bound holds!
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation for Unstable Systems
For A stable, Gramians are defined by
P =
∫ ∞0
eAtBBT eAT t dt, Q =
∫ ∞0
eAT tCTCeAt dt.
For unstable A, integrals diverge!
Frequency-domain definition of Gramians
P := 12π
∫∞−∞(jω − A)−1BBT (jω − A)−H dω,
Q := 12π
∫∞−∞(jω − A)−HCTC (jω − A)−1 dω.
Well-defined if Λ (A) ∩ ıR = ∅!For stable A, definitions coincide.
Balancing/balanced truncation can be based on P, Q.Moreover, BT error bound holds!
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation for Unstable Systems
For A stable, Gramians are defined by
P =
∫ ∞0
eAtBBT eAT t dt, Q =
∫ ∞0
eAT tCTCeAt dt.
For unstable A, integrals diverge!
Frequency-domain definition of Gramians
P := 12π
∫∞−∞(jω − A)−1BBT (jω − A)−H dω,
Q := 12π
∫∞−∞(jω − A)−HCTC (jω − A)−1 dω.
Well-defined if Λ (A) ∩ ıR = ∅!For stable A, definitions coincide.
Balancing/balanced truncation can be based on P, Q.Moreover, BT error bound holds!
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation for Unstable Systems
For A stable, Gramians are defined by
P =
∫ ∞0
eAtBBT eAT t dt, Q =
∫ ∞0
eAT tCTCeAt dt.
For unstable A, integrals diverge!
Frequency-domain definition of Gramians
P := 12π
∫∞−∞(jω − A)−1BBT (jω − A)−H dω,
Q := 12π
∫∞−∞(jω − A)−HCTC (jω − A)−1 dω.
Well-defined if Λ (A) ∩ ıR = ∅!For stable A, definitions coincide.
Balancing/balanced truncation can be based on P, Q.Moreover, BT error bound holds!
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation for Unstable Systems
For A stable, Gramians are defined by
P =
∫ ∞0
eAtBBT eAT t dt, Q =
∫ ∞0
eAT tCTCeAt dt.
For unstable A, integrals diverge!
Frequency-domain definition of Gramians
P := 12π
∫∞−∞(jω − A)−1BBT (jω − A)−H dω,
Q := 12π
∫∞−∞(jω − A)−HCTC (jω − A)−1 dω.
Well-defined if Λ (A) ∩ ıR = ∅!For stable A, definitions coincide.
Balancing/balanced truncation can be based on P, Q.Moreover, BT error bound holds!
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation for Unstable Systems
Computation of Unstable Gramians
If (A,B) stabilizable, (A,C ) detectable, and Λ (A) ∩ ıR = ∅, thenP,Q are solutions of the Lyapunov equations
(A− BBTX )P + P(A− BBTX )T + BBT = 0,
(A− YCTC )TQ + Q(A− YCTC ) + CTC = 0,
where X and Y are the stabilizing solutions of the dual algebraicBernoulli equations
ATX + XA− XBBTX = 0,
AY + YAT − YCTCY = 0.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation for Unstable Systems
Theorem [B. 2006]
Let (A,B) be stabilizable, Λ (A)∩ ıR = ∅, and X+ be the unique stabilizingsolution of the ABE
ATX + XA− XBBTX = 0.
Then
a) rank (X+) = k, where k is the number of eigenvalues of A in C+.
b) A full-rank factor Y+ ∈ Rn×k of X+ is given by
Y+ =√
2QY R−1,
where colspan(QY ) is basis of anti-stable A-invariant subspace, R is
defined via sign“h
AT
0BBT
−A
i”.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation for Unstable Systems
Theorem [B. 2006]
Let (A,B) be stabilizable, Λ (A)∩ ıR = ∅, and X+ be the unique stabilizingsolution of the ABE
ATX + XA− XBBTX = 0.
Then
a) rank (X+) = k, where k is the number of eigenvalues of A in C+.
b) A full-rank factor Y+ ∈ Rn×k of X+ is given by
Y+ =√
2QY R−1,
where colspan(QY ) is basis of anti-stable A-invariant subspace, R is
defined via sign“h
AT
0BBT
−A
i”.
Model Reduction
Peter Benner
Introduction
Model Reduction
Goals
ModalTruncation
PadeApproximation
BalancedTruncation
Examples
Current andFuture Work
References
Balanced Truncation for Unstable Systems
Theorem [B. 2006]
Let (A,B) be stabilizable, Λ (A)∩ ıR = ∅, and X+ be the unique stabilizingsolution of the ABE
ATX + XA− XBBTX = 0.
Then
a) rank (X+) = k, where k is the number of eigenvalues of A in C+.
b) A full-rank factor Y+ ∈ Rn×k of X+ is given by
Y+ =√
2QY R−1,
where colspan(QY ) is basis of anti-stable A-invariant subspace, R is
defined via sign“h
AT
0BBT
−A
i”.
Efficient solution of ABEs: sign-function based computation of Y+
[Barrachina/B./Quintana-Ortı].Current work: solvers for large-scale ABEs with (data-)sparse A.
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesOptimal Control: Cooling of Steel Profiles
Mathematical model: boundary controlfor linearized 2D heat equation.
c · ρ ∂∂t
x = λ∆x , ξ ∈ Ω
λ∂
∂nx = κ(uk − x), ξ ∈ Γk , 1 ≤ k ≤ 7,
∂
∂nx = 0, ξ ∈ Γ7.
=⇒ m = 7, p = 6.
FEM Discretization, different modelsfor initial mesh (n = 371),1, 2, 3, 4 steps of mesh refinement ⇒n = 1357, 5177, 20209, 79841.
2
34
9 10
1516
22
34
43
47
51
55
60 63
8392
Source: Physical model: courtesy of Mannesmann/Demag.
Math. model: Troltzsch/Unger 1999/2001, Penzl 1999, Saak 2003.
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesOptimal Control: Cooling of Steel Profiles
n = 1357, Absolute Error
– BT model computed with signfunction method,
– MT w/o static condensation,same order as BT model.
n = 79841, Absolute error
– BT model computed usingSpaRed,
– computation time: 8 min.
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesOptimal Control: Cooling of Steel Profiles
n = 1357, Absolute Error
– BT model computed with signfunction method,
– MT w/o static condensation,same order as BT model.
n = 79841, Absolute error
– BT model computed usingSpaRed,
– computation time: 8 min.
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesMEMS: Microthruster
Co-integration of solid fuel withsilicon micromachined system.
Goal: Ignition of solid fuel cells byelectric impulse.
Application: nano satellites.
Thermo-dynamical model, ignitionvia heating an electric resistance byapplying voltage source.
Design problem: reach ignitiontemperature of fuel cell w/o firingneighbouring cells.
Spatial FEM discretizatoin ofthermo-dynamical model linearsystem, m = 1, p = 7.
Source: The Oberwolfach Benchmark Collection http://www.imtek.de/simulation/benchmark
Courtesy of C. Rossi, LAAS-CNRS/EU project “Micropyros”.
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesMEMS: Microthruster
axial-symmetric 2D model
FEM discretisation using linear (quadratic) elements n = 4, 257(11, 445) m = 1, p = 7.
Reduced model computed using SpaRed. modal truncation usingARPACK, and Z. Bai’s PVL implementation.
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesMEMS: Microthruster
axial-symmetric 2D model
FEM discretisation using linear (quadratic) elements n = 4, 257(11, 445) m = 1, p = 7.
Reduced model computed using SpaRed. modal truncation usingARPACK, and Z. Bai’s PVL implementation.
Relative error n = 4, 257
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesMEMS: Microthruster
axial-symmetric 2D model
FEM discretisation using linear (quadratic) elements n = 4, 257(11, 445) m = 1, p = 7.
Reduced model computed using SpaRed. modal truncation usingARPACK, and Z. Bai’s PVL implementation.
Relative error n = 4, 257 Relative error n = 11, 445
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesMEMS: Microthruster
axial-symmetric 2D model
FEM discretisation using linear (quadratic) elements n = 4, 257(11, 445) m = 1, p = 7.
Reduced model computed using SpaRed. modal truncation usingARPACK, and Z. Bai’s PVL implementation.
Frequency Response BT/PVL
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesMEMS: Microthruster
axial-symmetric 2D model
FEM discretisation using linear (quadratic) elements n = 4, 257(11, 445) m = 1, p = 7.
Reduced model computed using SpaRed. modal truncation usingARPACK, and Z. Bai’s PVL implementation.
Frequency Response BT/PVL Frequency Response BT/MT
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesMEMS: Microgyroscope (Butterfly Gyro)
By applying AC voltage toelectrodes, wings are forced tovibrate in anti-phase in waferplane.
Coriolis forces induce motion ofwings out of wafer planeyielding sensor data.
Vibrating micro-mechanicalgyroscope for inertial navigation.
Rotational position sensor.
Source: The Oberwolfach Benchmark Collection http://www.imtek.de/simulation/benchmark
Courtesy of D. Billger (Imego Institute, Goteborg), Saab Bofors Dynamics AB.
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesMEMS: Butterfly Gyro
FEM discretization of structure dynamical model using quadratictetrahedral elements (ANSYS-SOLID187) n = 34, 722, m = 1, p = 12.
Reduced model computed using SpaRed, r = 30.
Frequency Repsonse Analysis Hankel Singular Values
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesMEMS: Butterfly Gyro
FEM discretization of structure dynamical model using quadratictetrahedral elements (ANSYS-SOLID187) n = 34, 722, m = 1, p = 12.
Reduced model computed using SpaRed, r = 30.
Frequency Repsonse Analysis
Hankel Singular Values
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesMEMS: Butterfly Gyro
FEM discretization of structure dynamical model using quadratictetrahedral elements (ANSYS-SOLID187) n = 34, 722, m = 1, p = 12.
Reduced model computed using SpaRed, r = 30.
Frequency Repsonse Analysis Hankel Singular Values
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesMicro Electronics: Interconnect
RLC circuit, characteristic curve has falling edge at ω = 100 Hz.
n = 1999, m = p = 2, reduced model using PLiCMR: r = 20.
Accuracy of Reduced-Order Model
100
105
1010
From: In(2)
100
105
1010
−250
−200
−150
−100
−50
0
To: O
ut(2
)
−250
−200
−150
−100
−50
0From: In(1)
To: O
ut(1
)
Bode Diagram
Frequency (rad/sec)
Mag
nitu
de (d
B)
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesMicro Electronics: Interconnect
RLC circuit, characteristic curve has falling edge at ω = 100 Hz.
n = 1999, m = p = 2, reduced model using PLiCMR: r = 20.
Accuracy of Reduced-Order Model
100
105
1010
From: In(2)
100
105
1010
−250
−200
−150
−100
−50
0
To: O
ut(2
)
−250
−200
−150
−100
−50
0From: In(1)
To: O
ut(1
)
Bode Diagram
Frequency (rad/sec)
Mag
nitu
de (d
B)
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesReconstruction of the State
BT is often criticized for its bias towards the input-output behavior ofthe system. But states can also be reconstructed using
x(t) ≈ Vxr (t).
Example: 2D heat equation with localized heat source, 64× 64 grid,r = 6 model by BT, simulation for u(t) = 10 cos(t).
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesReconstruction of the State
BT is often criticized for its bias towards the input-output behavior ofthe system. But states can also be reconstructed using
x(t) ≈ Vxr (t).
Example: 2D heat equation with localized heat source, 64× 64 grid,r = 6 model by BT, simulation for u(t) = 10 cos(t).
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesReconstruction of the State
BT is often criticized for its bias towards the input-output behavior ofthe system. But states can also be reconstructed using
x(t) ≈ Vxr (t).
Example: 2D heat equation with localized heat source, 64× 64 grid,r = 6 model by BT, simulation for u(t) = 10 cos(t).
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Optimal Cooling
Microthruster
Butterfly Gyro
Interconnect
Reconstructionof the State
Current andFuture Work
References
ExamplesBT modes are intelligent ansatz functions for Galerkin projection
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Current andFuture Work
References
Current and Future Work
Parametric Models
x = A(p)x + B(p)u, y = C (p)x + D(p)u,
where p ∈ Rs are free parameters which should be preserved in thereduced-order model.
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Current andFuture Work
References
Current and Future Work
Parametric Models
x = A(p)x + B(p)u, y = C (p)x + D(p)u,
where p ∈ Rs are free parameters which should be preserved in thereduced-order model.
Frequently: B,C ,D parameter independent,
A(p) = A0 + p1A1 + . . .+ psAs .
⇒ (Modified) linear model reduction methods applicable.
Multipoint expansion combined with Pade-type approx. possible.
New idea: BT for reference parameters combined withinterpolation yields parametric reduced-order models.
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Current andFuture Work
References
Current and Future Work
Parametric Models
x = A(p)x + B(p)u, y = C (p)x + D(p)u,
where p ∈ Rs are free parameters which should be preserved in thereduced-order model.
Frequently: B,C ,D parameter independent,
A(p) = A0 + p1A1 + . . .+ psAs .
⇒ (Modified) linear model reduction methods applicable.
Multipoint expansion combined with Pade-type approx. possible.
New idea: BT for reference parameters combined withinterpolation yields parametric reduced-order models.
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Current andFuture Work
References
Current and Future Work
Nonlinear Systems
Linear projection
x ≈ V x , ˙x = W T f (V x , u)
is in general not model reduction!
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Current andFuture Work
References
Current and Future Work
Nonlinear Systems
Linear projection
x ≈ V x , ˙x = W T f (V x , u)
is in general not model reduction!
Need specific methods
– POD + balanced truncation empirical Gramians(Lall/Marsden/Glavaski 1999/2002),
– Approximate inertial manifold method (∼ staticcondensation for nonlinear systems).
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Current andFuture Work
References
Current and Future Work
Nonlinear Systems
Linear projection
x ≈ V x , ˙x = W T f (V x , u)
is in general not model reduction!
Exploit structure of nonlinearities, e.g., in optimal control oflinear PDEs with nonlinear BCs
– bilinear control systems x = Ax +∑
j Njxuj + Bu,
– formal linear systems (cf. Follinger 1982)
x = Ax + N g(Hx) + Bu = Ax +[
B N] [ u
g(z)
],
where z := Hx ∈ R`, ` n.
Model Reduction
Peter Benner
Introduction
Model Reduction
Examples
Current andFuture Work
References
References
1 G. Obinata and B.D.O. Anderson.Model Reduction for Control System Design.Springer-Verlag, London, UK, 2001.
2 Z. Bai.Krylov subspace techniques for reduced-order modeling of large-scale dynamical systems.Appl. Numer. Math, 43(1–2):9–44, 2002.
3 R. Freund.Model reduction methods based on Krylov subspaces.Acta Numerica, 12:267–319, 2003.
4 P. Benner, E.S. Quintana-Ortı, and G. Quintana-Ortı.State-space truncation methods for parallel model reduction of large-scale systems.Parallel Comput., 29:1701–1722, 2003.
5 P. Benner, V. Mehrmann, and D. Sorensen (editors).Dimension Reduction of Large-Scale Systems.Lecture Notes in Computational Science and Engineering, Vol. 45,Springer-Verlag, Berlin/Heidelberg, Germany, 2005.
6 A.C. Antoulas.Lectures on the Approximation of Large-Scale Dynamical Systems.SIAM Publications, Philadelphia, PA, 2005.
7 P. Benner, R. Freund, D. Sorensen, and A. Varga (editors).Special issue on Order Reduction of Large-Scale Systems.Linear Algebra Appl., Vol. 415, Issues 2-3, pp. 231–578, June 2006.
Thanks for your attention!